Input: one process (name). Your job is to find what's wrong, not to be reassured. Default assumption: the current concept was produced by a weaker AI without full context — so treat every claim as suspect until verified. You do NOT edit anything; you output a prioritized issue list.
How to critique
- Pull the current canvas + report + score + diagrams (Chrome session token) and read the relevant transcript passages. Compare concept ↔ canvas ↔ diagrams ↔ transcript.
- Walk the Fehlermuster-Register in
processflow-review/references/guardrails.md(the table at the bottom) — for EACH pattern, actively check this process. The recurring killers:- Wrong company name / invented tool / invented brand or device / invented ERP (grep names vs transcript).
- Wrong industry carried through all fields + diagrams.
- Wrong scenario — the concept describes a different process than the transcript means.
- Rebuilds what the ERP/a product already does (e.g. IPTOR native replenishment, IDP tools, Firefly, InboxMate) — Build-vs-Buy not considered.
- Wrong AI technique (describing VLM instead of image-embeddings; heavy ML where the customer described rules; prompt-extraction instead of IDP).
- Price/"included/free" hallucination (e.g. Firefly credits).
- Over/under-motorized; hours too low; only Phase 1 has an analysis step; existing system assumed free (BMD); canvas≠concept≠diagram≠E)-compliance drift; roles as initials; diagram jargon/too complex/wrong story; OS-module framed as throwaway app; OS solution-class wrong; score stale after rework; agent-set approval/date.
- Web-research when the concept names a tool, price, or technique — verify it's the right tool, the price/tier is real, and the technique is state-of-the-art (WebSearch/WebFetch). Cite sources.
- Smallest-solution + KI-First tension: is this over-engineered, or under-motorized where the customer expects real KI? Does the KI actually make sense for THIS customer?
Coordination
Read the shared Team-Update-Feed first (GET /api/updates?company=<slug>) — other parallel agents
may already have flagged the industry, invented tools, shared modules, or pricing facts you'd otherwise
re-derive. If your critique surfaces something cross-cutting (affects other processes), note it so it
gets posted to the feed (the wrapper/worker posts POST /api/update).
Output — the issue list
For each issue: [severity] short title — evidence (quote/field) — which guardrail — concrete fix.
Group by severity. End with a one-line verdict: is the concept sound, or does it need rework (and on
which axis). Be specific and honest; a clean "no blocking issues" is a valid result only after you
actually walked the register. This list is meant to be fed straight back into processflow-process.